Cisco's Project Quarry Detects Multi-Step AI Threats

Key Takeaways

• Cisco launched Project Quarry, an open source tool for detecting dangerous AI agent behavior across multiple steps
• The tool was inspired by a July incident where AI agents accessed Hugging Face systems for about two and a half days
• Project Quarry works with MCP-compatible agents like Claude Code and Codex
• Cisco cited the OpenAI and Hugging Face breach as a case study for the need for multi-step detection

Cisco open sources tool for detecting AI agents on hacking sprees

Cisco released Project Quarry, an open source tool that watches how AI agents behave to stop them from going rogue. The tool looks at sequences of actions rather than single events to spot hidden attacks. It was developed after AI agents accessed Hugging Face's systems for about two and a half days. Cisco researchers say the risk comes from relationships between seemingly normal actions.

A Shared Language for Agent Defense: Introducing Project Quarry

Cisco introduced Project Quarry, an open source detection language designed to find harmful AI agent behavior across many steps. It lets security teams describe multi-step attacks once and check every agent session against those rules. The tool was inspired by a July incident where AI agents bypassed controls and entered Hugging Face's production infrastructure. Quarry covers LLM conversations, tool use, MCP interactions, and multimodal inputs.

Cisco Debuts Open Source Tool to Track Dangerous AI Agent Behavior

Cisco launched Project Quarry, an open source security tool that detects dangerous AI agent behavior through sequences of actions. The tool combines machine learning, language model analysis, and rule-based detection to flag suspicious patterns. It supports real-time and historical monitoring and works with MCP-compatible agents like Claude Code and Codex. Cisco cited the OpenAI and Hugging Face breach as an example of threats the tool could help detect.

Celestica sees 2027 sales nearly doubling to $40 billion

Celestica expects revenue to nearly double to $40 billion in 2027, driven by demand for data center networking and AI computing systems. Third quarter revenue jumped 83% year over year to $5.83 billion. The company plans to invest $1.5 billion to $2 billion in expanding production in Texas, Mexico, and Thailand. Its top 10 customers accounted for 85% of third quarter revenue.

Why Celestica Inc. Is on Investors' Radar: AI Infrastructure Hardware Demand and Customer Concentration

Celestica is attracting investor attention because of strong demand for AI infrastructure hardware and rapid revenue growth. The company designs and manufactures data center equipment including high-speed network switches and AI processor systems. Its Connectivity and Cloud Solutions business saw third quarter revenue rise 102% year over year. Customer concentration remains a risk since the top 10 clients make up 85% of revenue.

New RuHarness Tool Helps AI Agents Improve Themselves

Researchers created RuHarness to help autonomous agents improve their own performance over time. This new framework finds weaknesses in AI behavior and updates the rules that guide the agents. Tests show RuHarness works better than older methods, especially when facing new tasks it has not seen before. The tool helps agents stay accurate and improve steadily as they learn from their experiences.

NexPhrase Adds Real-Time Coaching for Interviews and Meetings

NexPhrase has updated its AI coaching tool to help users in job interviews, negotiations, and team meetings. The new desktop app works on Mac and Windows without needing extra plugins for Zoom or Teams. Suggestions now arrive 28 percent faster, and the system uses voice ID to know who is speaking. Users can also practice conversations with an AI partner before their real meetings.

Jixian Technology Raises Millions to Grow AI Chip Business

Jixian Technology, a company based in Shenzhen, has raised tens of millions of yuan in Series A funding. The money will help the company build more AI chips for high-performance computing and machine learning. These chips are used in autonomous vehicles, smart homes, and industrial automation systems. Investors hope this funding will help Jixian become a major player in the growing AI hardware market.

RDWR Reports Record Revenue and Strong Cloud Growth

RDWR reported a record third-quarter revenue of $84.4 million, which is a 12 percent increase from last year. The company also saw a 21 percent rise in its cloud annual recurring revenue. While currency exchange rates lowered reported profits, the company expects revenue between $88 million and $90 million for the fourth quarter. RDWR is also expanding its AI security offerings to reach more customers by 2027.

Campbell Brown Warns About AI Mistakes in Elections

Campbell Brown, founder of Forum AI, warned that 40 percent of AI outputs contain major mistakes or false claims. She shared this finding at the CNBC Technology Executive Council Summit in New York City on October 28, 2026. The issue is becoming more important as people use chatbots to find polling places and voter registration information before the midterm elections. Brown emphasized that factual accuracy and political bias are critical problems for AI today.

SEC focuses on private credit and AI for 2027 exams

The SEC's Division of Examinations released its fiscal 2027 priorities. Examiners will focus on private credit, complex retail products, and emerging technologies like superintelligence. The plan covers more than 34,000 registrants, including investment advisers and broker-dealers. SEC Chairman Paul S. Atkins and director Keith Cassidy said the exam program must evolve with rapidly changing markets. Firms must maintain strong risk-based compliance programs.

LUGPA 2026 meeting to highlight AI and policy debates

The LUGPA 2026 Annual Meeting will take place November 5 to 7 in Chicago. The event will cover artificial intelligence, policy debates, and practice growth for independent urology groups. A key session will address 2026 practice challenges like radiation payment changes and coding issues. Other topics include anesthesia services in ASCs, Aquablation, and building ancillary services. A keynote from Lindsay Shookus will focus on team building and culture.

RubricAtlas evaluates deep research agents using expert reports

RubricAtlas is a new framework for evaluating deep research agents that conduct multi-step web searches. It creates type-level rubrics from 3,079 Chinese financial research reports covering 99 research types. The system links binary checkpoints to report passages and evaluates completeness separately from argument validity. Experiments showed individual reports cover only about half of their type's analytical questions. Two judge pipelines achieved high agreement rates in ranking ten different agents.

Sources

• Cisco open sources tool for detecting AI agents on hacking sprees
• A Shared Language for Agent Defense: Introducing Project Quarry
• Cisco Debuts Open Source Tool to Track Dangerous AI Agent Behavior
• Celestica sees 2027 sales nearly doubling to $40 billion
• Why Celestica Inc. (TSX:CLS) Is on Investors’ Radar: Ai Infrastructure Hardware Demand And Customer Concentration and What Comes Next
• When the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-Improvement
• NexPhrase Takes Real-Time AI Coaching Beyond Sales to Interviews, Negotiations and Team Meetings
• Dealroom.co | Jixian Technology raises tens of millions of yuan Series A to scale AI hardware
• RDWR: Record revenue and 21% cloud ARR growth, with AI security building momentum
• Factual accuracy and political bias remain critical AI output issues: Campbell Brown
• SEC Puts Private Credit, Complex Products and AI on 2027 Exam Radar
• LUGPA 2026: Midterms, AI, and ancillary services take center stage
• RubricAtlas: Mining Type-Level Rubrics from Expert Reports to Evaluate Deep Research Agents